[openstack-dev] [Murano] SQLite support - drop or not?

Mike Bayer mbayer at redhat.com
Tue Feb 3 17:05:12 UTC 2015



Andrew Pashkin <apashkin at mirantis.com> wrote:

> Mike Bayer wrote:
>> there’s always a naming convention in place; all databases other than
>> SQLite produce them on the fly if you don’t specify one.  The purpose
>> of the Alembic/SQLAlchemy naming_convention feature is so that you
>> have *one* naming convention, rather than N unpredictable conventions.
>> I’m not sure if you’re arguing the feature should not be used.  IMHO
>> it should definitely be used for an application that is deploying
>> cross-database.  Otherwise you have no choice but to hardcode the
>> naming conventions of each target database individually in all cases
>> that you need to refer to them.
> You can't just bring SA/Alembic "naming conventions" into the project,
> because they will collide with auto-generated constraint names.

I was proposing a way to fix this for the murano project which only appears to have four migrations so far, but with the assumption that there are existing production environments which cannot do a full rebuild.

> 
> So you need to hardcode reverese-engineered constrants names into the
> old migrations and then add new migration that renames constraint
> according with "naming conventions”.
> OR you need to drop old
> migrations, and create new one with "naming conventions" - that will
> be backward incompatible, but cleaner.


My proposal was to essentially do both strategies.   Build out fully clean migrations from the start, but also add an additional “conditional” migration that will repair a Postgresql / MySQL database that is already at the head, and is detected as having the older naming convention.  Because openstack does not appear to use offline migrations, this would be doable, though not necessarily worth it.

If Murano can afford to just restart with clean migrations and has no production deployments yet which would be disrupted by a full rebuild, then sure, just do this.





> 
> On 03.02.2015 18:32, Mike Bayer wrote:
>> Andrew Pashkin <apashkin at mirantis.com> wrote:
>> 
>>> Mike Bayer wrote:
>>>> The patch seems to hardcode the conventions for MySQL and Postgresql.
>>>> The first thought I had was that in order to remove the dependence
>>>> on them here, you’d need to instead simply turn off the
>>>> “naming_convention” in the MetaData if you detect that you’re on one
>>>> of those two databases. That would be a safer idea than trying to
>>>> hardcode these conventions (and would also work for other kinds
>>>> of backends).
>>> With your solution it is still will be necessary for developers
>>> to guess constraints names when writing new migrations. And it will
>>> be even harder, because they will need also to handle case of
>>> "naming conventions”.
>> 
>> there’s always a naming convention in place; all databases other than SQLite produce them on the fly if you don’t specify one.  The purpose of the Alembic/SQLAlchemy naming_convention feature is so that you have *one* naming convention, rather than N unpredictable conventions.   I’m not sure if you’re arguing the feature should not be used.  IMHO it should definitely be used for an application that is deploying cross-database.  Otherwise you have no choice but to hardcode the naming conventions of each target database individually in all cases that you need to refer to them.
>> 
>> 
>> 
>> 
>>> Mike Bayer wrote:
>>>> However, it’s probably worthwhile to introduce a migration that does
>>>> in fact rename existing constraints on MySQL and Postgresql.
>>> Yes, that's what I want to do in case of the first solution.
>>> 
>>> Mike Bayer wrote:
>>>>> Another possible solution is to drop all current migrations and
>>>>> introduce new one with correct names.
>>>> you definitely shouldn’t need to do that.
>>> Why?
>>> 
>>> On 30.01.2015 22:00, Mike Bayer wrote:
>>>> Andrew Pashkin <apashkin at mirantis.com> wrote:
>>>> 
>>>>> Working on this issue I encountered another problem.
>>>>> 
>>>>> Most indices in the project has no names and because of that,
>>>>> developer must reverse-engineer them in every migration.
>>>>> Read about that also here [1].
>>>>> 
>>>>> SQLAlchemy and Alembic provide feature for generation constraint
>>>>> names by pattern, specifically to resolve that kind of issues [1].
>>>>> 
>>>>> I decided to introduce usage of this feature in Murano.
>>>>> 
>>>>> I've implemented solution that preserves backward-compatibility
>>>>> for migration and allows to rename all constraints according
>>>>> to patterns safely [2]. With it user, that have already deployed Murano
>>>>> will be able to upgrade to new version of Murano without issues.
>>>>> 
>>>>> There are downsides in this solution:
>>>>> - It assumes that all versions of Postgres and MySQL uses the
>>>>> same patterns for constraints names generation.
>>>>> - It is hard to implement a test for this solution and it will be slow.
>>>>> Because there is need to reproduce such situation when user has old
>>>>> versions of migrations applied, and then tries to upgrade.
>>>> 
>>>> The patch seems to hardcode the conventions for MySQL and Postgresql.   The first thought I had was that in order to remove the dependence on them here, you’d need to instead simply turn off the “naming_convention” in the MetaData if you detect that you’re on one of those two databases.   That would be a safer idea than trying to hardcode these conventions (and would also work for other kinds of backends).
>>>> 
>>>> However, I’m not actually sure that you even need special behavior for these two backends.  If an operator runs these migrations on a clean database, then the constraints are generated with the consistent names on all backends.   if a target database already has these schema constructs present, then these migrations are never run; it doesn’t matter that they have the right or wrong names already.
>>>> 
>>>> I suppose then that the fear is that some PG/MySQL databases will have constraints that are named in one convention, and others will have constraints using the native conventions.    However, the case now is that all deployments are using native conventions, and being able to DROP these constraints is already not very feasible unless you again were willing to hardcode those naming conventions up forward.    The constraints in these initial migrations, assuming you don’t regenerate them, might just need to be left alone, and the project proceeds in the future with a consistent convention.
>>>> 
>>>> However, it’s probably worthwhile to introduce a migration that does in fact rename existing constraints on MySQL and Postgresql.  This would be a migration script that emits DROP CONSTRAINT and CREATE CONSTRAINT for all the above constraints that have an old name and a new name.  The script would need to check the backend, as you’re doing now, in order to run, and yes it would hardcode the names of those conventions, but at least it would just be a one-time run against only currently deployed databases.   Since your migrations are run “live”, the script can make itself a “conditional” run by checking for the “old” names and skipping those that don’t exist.      
>>>> 
>>>>> Another possible solution is to drop all current migrations and
>>>>> introduce new one with correct names.
>>>> 
>>>> you definitely shouldn’t need to do that.
>>>> 
>>>> 
>>>>> This brings us to new problem - migrations and models are out of sync
>>>>> right now in multiple places - there are different field types in
>>>>> migrations and models, migrations introduces indices that is absent
>>>>> in models, etc.
>>>>> 
>>>>> And this solution has great downside - it is not backward-compatible,
>>>>> so all old users will lost their data.
>>>>> 
>>>>> We (Murano team) should decide, what solution we want to use.
>>>>> 
>>>>> 
>>>>> [1]
>>>>> http://alembic.readthedocs.org/en/latest/naming.html#tutorial-constraint-names
>>>>> [2] https://review.openstack.org/150818
>>>>> 
>>>>> -- 
>>>>> With kind regards, Andrew Pashkin.
>>>>> cell phone - +7 (985) 898 57 59
>>>>> Skype - waves_in_fluids
>>>>> e-mail - apashkin at mirantis.com
>>>>> 
>>>>> __________________________________________________________________________
>>>>> OpenStack Development Mailing List (not for usage questions)
>>>>> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>> 
>>>> __________________________________________________________________________
>>>> OpenStack Development Mailing List (not for usage questions)
>>>> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>> 
>>> -- 
>>> With kind regards, Andrew Pashkin.
>>> cell phone - +7 (985) 898 57 59
>>> Skype - waves_in_fluids
>>> e-mail - apashkin at mirantis.com
>>> 
>>> __________________________________________________________________________
>>> OpenStack Development Mailing List (not for usage questions)
>>> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>> 
>> __________________________________________________________________________
>> OpenStack Development Mailing List (not for usage questions)
>> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> 
> -- 
> With kind regards, Andrew Pashkin.
> cell phone - +7 (985) 898 57 59
> Skype - waves_in_fluids
> e-mail - apashkin at mirantis.com
> 
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev



More information about the OpenStack-dev mailing list